Handcrafted Radiomics Features

• Handcrafted radiomics features are mathematically defined, interpretable descriptors extracted from segmented imaging regions to quantify intensity, shape, and texture.
• They rely on standardized preprocessing and parameterization protocols to ensure reproducibility and mitigate variability caused by scanner or segmentation differences.
• HRFs serve as a bridge to deep learning by enabling hybrid models that fuse engineered and learned features, enhancing performance on limited and heterogeneous datasets.

Handcrafted radiomics features (HRFs), also called engineered features, are quantitative descriptors computed from predefined image-derived measurements extracted from a segmented region or volume of interest. They convert imaging data into structured, high-dimensional feature representations while retaining closed-form mathematical definitions for intensity histograms, shape indices, texture descriptors, filter responses, and related biomarkers. In contrast to deep radiomics, HRFs are user-driven, mathematically explicit, and typically more interpretable, but their reliability depends on harmonized acquisition, segmentation, preprocessing, and validation protocols (Neha et al., 24 Jan 2026, Afshar et al., 2018, Lavrova et al., 2024).

1. Definition and conceptual position

In the standardized radiomics pipeline, HRFs fall into three major families—first-order statistics, shape descriptors, and texture features—each encoding distinct aspects of image data. First-order statistics summarize the intensity distribution inside the ROI without regard to spatial arrangement; shape descriptors quantify geometry independent of intensity; texture features encode spatial relationships between gray levels through statistical matrices such as the gray-level co-occurrence matrix (GLCM). Broader implementations additionally include higher-order features derived from wavelets, Laplacian-of-Gaussian or Gabor filtering, as well as moment-invariant and diagnostic features in software frameworks such as PySERA (Neha et al., 24 Jan 2026, Afshar et al., 2018, Salmanpour et al., 20 Nov 2025).

HRFs occupy a distinct methodological position relative to deep radiomics. Handcrafted radiomics is a user-driven workflow in which an ROI is defined, a fixed set of mathematically defined features is computed, and feature selection and modeling follow; deep radiomics embeds feature extraction, selection, and modeling in one or more neural-network layers. The trade-off reported across reviews is consistent: HRFs are transparent, interpretable, and workable with smaller datasets, whereas deep radiomics can capture more complex patterns but demands larger training sets and yields less explicit feature semantics (Lavrova et al., 2024, Afshar et al., 2018).

A common misconception is that mathematical explicitness alone guarantees robustness. The survey literature instead emphasizes that reliable HRF computation hinges on preprocessing choices and parameter settings, and that scanner- and segmentation-induced variability can overwhelm true biological signal when protocols are not standardized (Neha et al., 24 Jan 2026). This suggests that the defining property of HRFs is not simply interpretability, but interpretability under controlled extraction conditions.

2. Feature families and mathematical structure

First-order HRFs probe the intensity histogram of the ROI. For voxel intensities $\{x_i\}_{i=1}^N$, representative descriptors include the mean $\mu=\frac{1}{N}\sum_{i=1}^N x_i$, the variance $\sigma^2=\frac{1}{N}\sum_{i=1}^N (x_i-\mu)^2$, skewness, kurtosis, entropy, energy, uniformity, percentiles, and related summary statistics. Across reviews and application papers, these measures are used to characterize attenuation, signal intensity, enhancement behavior, and heterogeneity at the histogram level (Neha et al., 24 Jan 2026, Afshar et al., 2018).

Shape descriptors quantify lesion or organ geometry independently of gray-level intensity. Frequently used examples are volume, surface area, compactness, sphericity, surface-to-volume ratio, maximum diameter, and axis lengths. A representative sphericity definition is

$\Psi=\frac{\pi^{1/3}(6V)^{2/3}}{A},$

where $V$ is volume and $A$ is surface area. Such features are explicitly linked in the literature to irregularity, deviation from spherical morphology, and margin complexity (Na et al., 11 Jul 2025, Neha et al., 24 Jan 2026).

Texture HRFs model spatial relationships between gray levels. For a normalized GLCM $P(i,j)$, common statistics include contrast,

$\mathrm{Contrast}=\sum_{i,j}(i-j)^2P(i,j),$

correlation, energy $\sum_{i,j}P(i,j)^2$, homogeneity, dissimilarity, and entropy. Other matrix families extend the same logic to different spatial organizations: GLRLM measures runs of identical gray levels, GLSZM measures connected zones, NGTDM measures local gray-tone differences, and GLDM or NGLDM measure dependence structure (Neha et al., 24 Jan 2026, Afshar et al., 2018, Salmanpour et al., 20 Nov 2025).

Recent task-specific papers also extend the classical family structure. RadiomicNet uses Gray-Level Co-occurrence Matrix and Local Binary Pattern features, with LBP defined on a circular neighborhood by

$$\mathrm{LBP}_{P,R}(x_c)=\sum_{p=0}^{P-1}s(I(x_p)-I(x_c))\,2^p,\qquad s(u)= \begin{cases} 1,&u\ge 0\ 0,&u<0 \end{cases}$$

using $\mu=\frac{1}{N}\sum_{i=1}^N x_i$0 and $\mu=\frac{1}{N}\sum_{i=1}^N x_i$1; its 13-dimensional radiomics vector comprises 5 GLCM, 4 LBP, and 4 first-order intensity moments (Rahman, 2 Jul 2026). A different extension is Image Frequency Analysis, which computes Pearson-style correlations between adjacent rows or columns and then applies a one-dimensional Fourier transform to the resulting sequences, yielding a directional frequency descriptor of the image (Usuzaki et al., 2021). This suggests that HRFs are not limited to conventional ROI-level histogram and texture matrices, but also include explicitly designed descriptors of cross-line structure and local coding patterns.

3. Extraction pipeline, preprocessing, and parameterization

The canonical handcrafted-radiomics workflow comprises image acquisition and reconstruction, segmentation, preprocessing, feature extraction, and feature reduction or model building. Segmentation may be manual, semi-automatic, or fully automatic, and segmentation quality directly impacts feature stability. The preprocessing stage typically includes intensity normalization, spatial resampling to isotropic voxels, grayscale discretization, and, when required, denoising or artifact correction (Afshar et al., 2018, Neha et al., 24 Jan 2026).

Parameterization is not ancillary; it changes the feature being computed. Standard choices reported across surveys include spatial resampling to isotropic voxels such as $\mu=\frac{1}{N}\sum_{i=1}^N x_i$2, normalization by z-score, histogram matching, or Nyúl’s mapping, and gray-level discretization by either fixed bin width or fixed bin count. For texture matrices, the distance $\mu=\frac{1}{N}\sum_{i=1}^N x_i$3, directions $\mu=\frac{1}{N}\sum_{i=1}^N x_i$4, and aggregation strategy are all consequential: longer distances capture coarser patterns, while multi-directional averaging improves rotational invariance at the cost of interpretability (Neha et al., 24 Jan 2026, Afshar et al., 2018).

Concrete implementations make these dependencies explicit. RadiomicNet computes GLCM features over distances $\mu=\frac{1}{N}\sum_{i=1}^N x_i$5 and angles $\mu=\frac{1}{N}\sum_{i=1}^N x_i$6 with 64 gray-level quantization, and averages the five statistics over all $\mu=\frac{1}{N}\sum_{i=1}^N x_i$7 pairs; its LBP uses radius $\mu=\frac{1}{N}\sum_{i=1}^N x_i$8, $\mu=\frac{1}{N}\sum_{i=1}^N x_i$9 neighbors, and uniform encoding (Rahman, 2 Jul 2026). In PySERA, standardized preprocessing includes resampling, discretization, normalization, optional range-based re-segmentation, optional outlier removal, and partial-volume binarization; the framework computes 557 handcrafted features, including 487 IBSI-compliant features, 10 moment-invariant descriptors, and 60 diagnostic features (Salmanpour et al., 20 Nov 2025).

Software ecosystems formalize these workflows. The survey literature explicitly names PyRadiomics, IBEX, and CERR as radiomics software environments aligned with IBSI definitions, while PySERA extends this landscape with a Python-native, object-oriented implementation benchmarked against PyRadiomics, MITK, and LIFEx (Neha et al., 24 Jan 2026, Salmanpour et al., 20 Nov 2025). In practice, the software choice is inseparable from the definition of the extracted feature set, because implementation details, defaults, and compliance level affect reproducibility.

4. Reproducibility, stability, and dimensionality control

Feature instability is one of the central methodological problems in handcrafted radiomics. Major sources of variability include scanner manufacturer, reconstruction kernel, slice thickness, and ROI delineation; segmentation uncertainty propagates directly into feature values. The survey literature identifies two standard reproducibility measures: the intraclass correlation coefficient (ICC), with features above $\sigma^2=\frac{1}{N}\sum_{i=1}^N (x_i-\mu)^2$0 deemed reproducible, and the coefficient of variation (CoV) (Neha et al., 24 Jan 2026).

Harmonization is therefore a structural component of HRF analysis rather than a post hoc convenience. Feature-domain methods such as ComBat and Bayesian ComBat are reported as ways to remove batch effects while retaining biological variance, and IBSI provides reference definitions and test-case images to standardize implementation across software. At the same time, modality-specific results show that no single harmonization method is uniformly optimal. In an echocardiography study on left-ventricle hypertension disease phenotyping, ComBat yielded mean Jensen–Shannon divergence $\sigma^2=\frac{1}{N}\sum_{i=1}^N (x_i-\mu)^2$1 versus $\sigma^2=\frac{1}{N}\sum_{i=1}^N (x_i-\mu)^2$2 on original images, whereas a self-supervised convolutional-kernel filter reduced it to $\sigma^2=\frac{1}{N}\sum_{i=1}^N (x_i-\mu)^2$3 and increased AUC from $\sigma^2=\frac{1}{N}\sum_{i=1}^N (x_i-\mu)^2$4 to $\sigma^2=\frac{1}{N}\sum_{i=1}^N (x_i-\mu)^2$5 (Lee et al., 2023).

Because HRF studies often operate in the $\sigma^2=\frac{1}{N}\sum_{i=1}^N (x_i-\mu)^2$6 regime, feature selection and dimensionality reduction are equally central. Reported strategies include filter methods such as variance thresholding, correlation analysis, mutual information, and mRMR; wrapper methods such as recursive feature elimination; embedded methods such as LASSO and Elastic Net; and dimensionality reduction by PCA, kernel PCA, or autoencoders. A recurring methodological requirement is that all selection and reduction steps must reside inside nested cross-validation loops to avoid data leakage and optimistic bias (Neha et al., 24 Jan 2026).

The instability induced by multicollinearity has motivated dedicated statistical frameworks. One example is the four-step approach of redundancy filtering, penalized maximum-likelihood estimation of a filtered correlation matrix, maximum-likelihood factor analysis, and factor scoring, producing a compact latent-feature space for regression-based prediction. This approach was explicitly proposed to address the high dimensionality, extreme multicollinearity, and instability of standard penalized classifiers in radiomics data (Peeters et al., 2019). A plausible implication is that stable HRF modeling often depends as much on covariance structure control as on the semantic content of the original descriptors.

5. Hybridization with deep learning and task-specific deployment

A major recent development is the direct integration of HRFs into deep architectures rather than their use as separate tabular covariates. RadiomicNet is a two-stream hybrid segmentation network in which a radiomics embedding modulates skip-connection attention through a Radiomics Attention Gate, and a Radiomics Consistency Loss aligns texture complexity with prediction entropy. With 13 handcrafted radiomics features as input to the radiomics branch, the model achieved a Dice Similarity Coefficient of $\sigma^2=\frac{1}{N}\sum_{i=1}^N (x_i-\mu)^2$7 on BUSI and $\sigma^2=\frac{1}{N}\sum_{i=1}^N (x_i-\mu)^2$8 on Kvasir-SEG, reduced Expected Calibration Error on BUSI from $\sigma^2=\frac{1}{N}\sum_{i=1}^N (x_i-\mu)^2$9 to $\Psi=\frac{\pi^{1/3}(6V)^{2/3}}{A},$0, and used only $\Psi=\frac{\pi^{1/3}(6V)^{2/3}}{A},$1M parameters. Gradient-based feature importance identified GLCM dissimilarity ($\Psi=\frac{\pi^{1/3}(6V)^{2/3}}{A},$2), GLCM energy ($\Psi=\frac{\pi^{1/3}(6V)^{2/3}}{A},$3), and LBP entropy ($\Psi=\frac{\pi^{1/3}(6V)^{2/3}}{A},$4) as the dominant radiomics cues (Rahman, 2 Jul 2026).

Hybridization can also occur at both global and voxel levels. In pancreatic ductal adenocarcinoma detection, a two-stage nnUNet-based workflow first extracted 1,486 global radiomic descriptors over the pancreas mask, then selected ten features by a Pearson correlation test with false discovery rate control and recursive feature elimination wrapped around a linear SVM, and finally injected eight non-shape features as voxel-level parametric maps while using the ten-dimensional global vector in radiomics-aware cross-attention at the bottleneck. The resulting method achieved AUC $\Psi=\frac{\pi^{1/3}(6V)^{2/3}}{A},$5 and AP $\Psi=\frac{\pi^{1/3}(6V)^{2/3}}{A},$6 in cross-validation on PANORAMA, and AUC $\Psi=\frac{\pi^{1/3}(6V)^{2/3}}{A},$7 and AP $\Psi=\frac{\pi^{1/3}(6V)^{2/3}}{A},$8 on an external in-house cohort (Deng et al., 20 Feb 2026).

Patch-level fusion has been developed for transformer architectures. RE-ViT extracts 91 patch-level HRFs from six families and fuses them by early fusion with linearly projected image patches before ViT encoding. Reported results were AUC $\Psi=\frac{\pi^{1/3}(6V)^{2/3}}{A},$9 on BUSI, $V$0 on ChestXray2017, and $V$1 on Retinal OCT, with statistically significant degradation when the radiomics stream was removed (Yang et al., 15 Apr 2025).

Not all effective HRF deployments require large generic feature banks. In hepatocellular carcinoma classification on 3D CT, three handcrafted radiological features—arterial-phase hyper-enhancement, non-peripheral wash-out, and enhancing capsule—were derived from lesion and parenchyma measurements motivated by LI-RADS criteria and fused with deep outputs in an $V$2-regularized logistic regression. On the transfer set $V$3, HF + size alone achieved $V$4 AUC and the combined DLF + HF + size model achieved $V$5, outperforming the baseline “Tiny” net at $V$6 (Sarfati et al., 14 Jan 2025).

Prognostic and radiogenomic settings show a more mixed interaction between handcrafted and learned representations. In lung-cancer survival prediction, PET_HRF with semi-supervised learning reduced MAE from $V$7 years to $V$8 years, and CT_HRF with PCA plus Component-Wise Gradient Boosting Survival Analysis achieved external c-index $V$9 (Gorji et al., 31 Jan 2025). In multicenter NSCLC radiogenomics for EGFR/KRAS prediction, IBSI-compliant HRF-only pipelines generalized best, reaching external AUC $A$0 and accuracy $A$1, whereas fusion improved robustness over DFR-only models but did not consistently outperform HRFs (Mehrnia et al., 26 Mar 2026). In multicenter NSCLC survival modeling, ComBat-harmonized tumor radiomics combined with clinical covariates achieved C-index $A$2 and 5-year t-AUC $A$3 (Mali et al., 23 May 2025).

HRFs have also been extended to retrieval and generation. RadiomicsRetrieval uses 72 tumor-level HRFs aligned with promptable image embeddings and anatomical positional embedding; radiomics-based retrieval correlations reached $A$4 on BraTS and $A$5 on NSCLC for Top-5 HRF queries (Na et al., 11 Jul 2025). Radiomics-conditioned tumor synthesis uses 16 shape features and 74 texture features as conditioning vectors in a GAN-based shape generator and a diffusion-based texture generator, while RadiomicsFill conditions glioma synthesis on 67 HRFs to control tumor geometry and heterogeneity (Kim et al., 29 Sep 2025, Na et al., 2023).

6. Interpretability, misconceptions, and open problems

The principal interpretive appeal of HRFs is that downstream decisions can be traced to named image phenotypes. RadiomicNet describes its attention modulation as a direct, differentiable function of named radiomics features with clear clinical meaning, and its learned scalar gate and channel weights are tied to specific texture descriptors (Rahman, 2 Jul 2026). Patient-specific radiomic feature selection for knee MRI retains fully interpretable radiomic features while learning a sparse per-patient subset, with case studies linking top-weighted features such as entropy, interquartile range, compactness, energy, uniformity, and the 10th percentile to clinically meaningful anatomy and pathology (Chen et al., 17 Mar 2025). In multicenter CT radiogenomics, SHAP highlighted morphology- and heterogeneity-related radiomic phenotypes such as Morph-VolDens-AABB, coefficient of variation, moment invariants, GLDZM-SDHGE, and GLRLM-LRE as key predictors (Mehrnia et al., 26 Mar 2026).

Interpretability, however, should not be conflated with automatic validity. Reviews emphasize persistent challenges of feature instability, limited reproducibility, validation bias, domain shift, and restricted clinical translation. The survey literature calls for harmonized protocols, unbiased validation frameworks, standardized multi-center benchmarks, federated learning, multimodal fusion, and regulatory-compliant, end-to-end reproducible workflows with uncertainty quantification (Neha et al., 24 Jan 2026). The PET/SPECT review similarly reports that only $A$6 of studies explicitly adhered to IBSI definitions, that class imbalance was addressed in only $A$7, that missing data were discussed in $A$8, and that low population diversity remained common; across 226 studies, deep radiomics achieved the highest mean accuracy ($A$9) while fusion yielded the highest mean AUC ($P(i,j)$0), indicating that handcrafted radiomics is neither obsolete nor universally dominant (Salmanpour et al., 21 Jul 2025).

A second misconception is that fusion with deep learning invariably improves performance. The empirical record is more conditional. Hybrid methods can be highly effective when HRFs inject heterogeneity priors, spatially resolved parametric maps, or clinically grounded descriptors into the network, but multicenter radiogenomics results show that HRF-only models may still generalize better than DFR-only or fused systems under strong site variation (Deng et al., 20 Feb 2026, Mehrnia et al., 26 Mar 2026). This suggests that the enduring value of HRFs lies less in competing with learned features on representational capacity than in providing stable, inspectable, and domain-grounded structure for modeling under limited data, heterogeneous acquisition, and translational constraints.